wtf is qwen-native-agents?

This repository is a guide and toolkit for building AI agents that work with the Qwen3.x family of language models, using the exact conventions those models were trained on rather than generic best practices. The core idea is that Qwen models learned their habits alongside two specific frameworks built by Alibaba, called qwen-code and Qwen-Agent. The way those frameworks format tool calls, structure system prompts, and handle the model's internal reasoning steps is baked into the model's training, not just a style choice. The project argues that if you want a Qwen agent to behave reliably, you should speak to it in that native dialect instead of trying to retrain its habits with your own prompt style. The repository is organized as a skill file plus a set of reference documents. The main file, SKILL.md, is a short guide covering when to use these conventions and a checklist of rules. A references folder breaks the topic into focused documents: how to write a system prompt in the expected tone, how to name and describe tools, the different formats models use to call tools, how the thinking or reasoning mode works across different backends, and how special tags inside prompts should be structured. There are also example files showing a complete system prompt, a tool definition, and a minimal agent setup. A person can read SKILL.md as a quick reference, while an AI coding agent like Claude Code can load the folder as a skill and pull it in automatically when a task involves building a Qwen agent. To support its claims, the project includes benchmark experiments comparing Qwen models against other open models like gpt-oss-120b and gemma, measuring things like how strongly each model sticks to habits such as writing few code comments, using little markdown, or calling tools in a specific format, even when instructed otherwise. The benchmarks also test which reasoning mode setting works best for each model and which tool-calling format each one performs best with. The raw benchmark code and result data are included in the repository for anyone who wants to reproduce or extend the tests. This project is aimed at developers building AI agents on top of Qwen models who want more predictable, reliable behavior by matching the model's trained conventions rather than fighting against them.